Model-based surface soil moisture (SSM) retrieval algorithm using multi-temporal RISAT-1 C-band SAR data

Abstract

Accurate measurement of surface soil moisture of bare and vegetation covered soil over agricultural field and monitoring the changes in surface soil moisture is vital for estimation for managing and mitigating risk to agricultural crop, which requires information and knowledge to assess risk potential and implement risk reduction strategies and deliver essential responses. The empirical and semi-empirical model-based soil moisture inversion approach developed in the past are either sensor or region specific, vegetation type specific or have limited validity range, and have limited scope to explain physical scattering processes. Hence, there is need for more robust, physical polarimetric radar backscatter model-based retrieval methods, which are sensor and location independent and have wide range of validity over soil properties. In the present study, Integral Equation Model (IEM) and Vector Radiative Transfer (VRT) model were used to simulate averaged backscatter coefficients in various soil moisture (dry, moist and wet soil), soil roughness (smooth to very rough) and crop conditions (low to high vegetation water contents) over selected regions of Gujarat state of India and the results were compared with multi-temporal Radar Imaging Satellite-1 (RISAT-1) C-band Synthetic Aperture Radar (SAR) data in σ° HH and σ° HV polarizations, in sync with on field measured soil and crop conditions. High correlations were observed between RISAT-1 HH and HV with model simulated σ° HH & σ° HV based on field measured soil with the coefficient of determination R 2 varying from 0.84 to 0.77 and RMSE varying from 0.94 dB to 2.1 dB for bare soil. Whereas in case of winter wheat crop, coefficient of determination R 2 varying from 0.84 to 0.79 and RMSE varying from 0.87 dB to 1.34 dB, corresponding to with vegetation water content values up to 3.4 kg/m 2 . Artificial Neural Network (ANN) methods were adopted for model-based soil moisture inversion. The training datasets for the NNs were obtained from theoretical forward-scattering models with controlled parameters, thus allowing the control of wide range of soil and crop parameters with which the network was trained. A preliminary performance analysis showed good results with estimation of soil moisture with RMSE better than 6%.